I am working with dynamic Bayesian networks with a mixture
of continuous and discrete variables. I am using the
propagation algorithm given in Lauritzen and Jensen (2001),
"Stable local computation with conditional Gaussian distributions,"
Statistics and Computing.
Consider the following DBN, with discrete variables D[1] . . . D[N]
and continous variables C[1] . . . C[N], where any variable can be
unobserved:
D[1] ---> D[2] ---> D[3] ---> D[4] ---> . . . ---> D[N]
| | | | |
| | | | |
| | | | |
| | | | |
V V V V V
C[1] ---> C[2] ---> C[3] ---> C[4] ---> . . . ---> C[N]
Now, in order to have a decomposable graph after triangulation,
there can not be any path between two non-adjacent discrete variables
that passes only through continuous vertices (Proposition 7.9
of Cowell et al., (1999) "Probabilistic Networks and Expert Systems).
In this case, this means that all the discrete vertices D[1] . . . D[N]
must be fully connected (because, since there is such a path between
every possible pair of discrete vertices, the only solution is to
connect them all with each other in triangulation).
============================================================
My question is the following:
Is there a method to avoid having to fully connect all these
discrete variables and thus have to have a large clique containing
all of the discrete variables? I would prefer, rather, to have
smaller cliques (e.g. the clique {D[t],D[t+1],C[t],C[t+1]}) when
doing propagation.
Thanks in advance for any input.
Todd
-- Todd Stephenson Dalle Molle Institute for Perceptual Artificial Intelligence PO Box 592 CH-1920 Martigny Switzerland Tel: +41 27 721 77 52 Fax: +41 27 721 77 12 Email: Todd.Stephenson@idiap.ch WWW: http://www.idiap.ch
This archive was generated by hypermail 2b29 : Tue Nov 27 2001 - 10:28:26 PST